G protein–coupled receptors (GPCRs) are among the most important drug targets in human biology, but developing antibodies that can not only bind GPCRs but also modulate their function – particularly as agonists – remains challenging. Most antibody discovery workflows prioritize binding affinity, leaving the harder question of pharmacological activity until much later. For antagonists this can be sufficient, but for agonists, function is inseparable from structure, orientation, and dynamics.
Abalone Bio is attempting to flip the discovery process by screening vast antibody libraries for function first, and then refining hits with the help of AI-driven protein language models. The company’s lead programs focus on cannabinoid receptor 2 (CB2), a GPCR expressed on immune cells and implicated in fibrosis, inflammation, and neuropathic pain. Unlike its better-known cousin CB1, which is concentrated in the central nervous system, CB2 is a peripheral target, which makes it attractive for interventions that modulate immune-driven pathology without psychoactive effects.
Together with collaborators at the Icahn School of Medicine at Mount Sinai, Abalone Bio has published preclinical results on CB2 agonist antibodies that reduce inflammation and fibrotic signaling in human liver tissue and show efficacy in animal models of neuropathy. For diseases such as liver fibrosis and chronic neuropathic pain, where current treatments are limited and often come with severe side effects, the prospect of a highly specific antibody agonist offers a fundamentally different path forward.
In this interview, Abalone Bio CEO Richard Yu discusses the challenges of functional antibody discovery, why CB2 is their first target, and how the company is using AI to accelerate the path from discovery to clinic.
What major challenges tend to hinder the discovery of functional antibody agonists?
Fundamentally, it’s a focus on function. Traditional antibody discovery looks for the strongest binders first, and only later assesses whether those high affinity molecules actually affect the function of the target. This approach works fine for antagonists that often work by binding in a way that prevents interaction with a ligand. But for agonists, it’s about the specific orientation and dynamics of binding, and how that binding affects the activity – not just the affinity of the binding event. Identifying binding interactions that are pharmacologically active –especially agonists – is particularly challenging for GPCRs. These are large, complex, and dynamic membrane proteins that are difficult to study using both wet-lab and computational methods.
We flipped that process by testing for function first, across hundreds of millions of antibody variants, by harnessing a fundamental tool from biological evolution – survival of the fittest. We engineered yeast cells to express our GPCR target on the cell membrane surface and linked its activity to the cell’s ability to grow. Further engineering each yeast cell resulted in a unique antibody variant that interacts with the receptor outside the cell, just like in the patient context. If the antibody activates the receptor, the cell grows faster. By keeping track of which cells grow faster by next-generation sequencing, we can identify pharmacologically active antibody sequences. That sequence-function data is then fed into large language models, which can improve signal-to-noise and better identify active candidates, as well as design novel antibody sequences not present in the original library.
For many GPCRs, where there is paucity of structural data and structural prediction methods fall short in capturing the dynamic components of function, direct functional data can fill the gap.
Why did you choose CB2 as your first target?
Our strategy is to pursue targets where previous drug approaches have run into limitations. We believe pharmacologically active antibodies can address many of these gaps, offering specificity, long half-lives, exclusion from the CNS, and the flexibility to be engineered into bispecific or multi-domain formats for targeted and multifunctional therapies. CB2 appealed to us because of preclinical evidence indicating its involvement in multiple areas that underlie diseases such as fibrosis, inflammation, and pain.
GPCRs are involved in practically every area of biology, so there are plenty of places to make a positive impact. That said, we’re first focusing on validated targets in metabolism, immunology, and inflammation, for which antibodies are the best modality to achieve the target product profile to most effectively treat the disease for the largest number of patients.
How do your CB2-targeting antibodies offer advantages over previous small-molecule drugs aimed at the same receptor?
Antibodies have several advantages when it comes to CB2 specificity over CB1. The extracellular regions of CB2 that an antibody would recognize are highly different from those of CB1, with only about 26 percent sequence homology. That makes it very unlikely for an antibody to bind both receptors, let alone to bind in a way that changes the structural dynamics needed to activate signaling in both. In addition, CB1 is mostly expressed in the central nervous system, and antibodies generally cannot cross the blood–brain barrier—so they are physically prevented from reaching CB1 in the first place.
How confident are you that your fibrosis and neuropathy models will translate clinically?
It’s no secret that these two disease areas have historically been quite challenging to develop therapies for, despite great unmet medical needs. Translation from animal results to humans is generally fraught, but we are optimistic for successful clinical translation based on the biology underlying the diseases and some examples of preclinical to clinical efficacy translation
Drugs for liver fibrosis and primary biliary cholangitis using rodent models of liver disease have translated to human efficacy and approval. Of course, rodent models of liver disease are not perfect representations of human disease biology. We have tried to bridge that gap in part by confirming anti-inflammatory activities in precision-cut human liver slices. In the context of human tissue, our antibodies reduce pro-fibrotic and pro-inflammatory gene expression, giving us confidence that we are not observing mere mouse- or rat-specific activities. From the fundamental biology angle, there is also supporting human genetic evidence regarding the role of CB2 in liver inflammation, including observations that hepatitis C patients with partial loss-of-function CB2 genetic variants have more severe hepatic inflammation and necrosis, supporting hepatic anti-inflammatory activities of CB2.
For peripheral neuropathy, the biology supports clinical translation. Macrophage-mediated neuroinflammation contributes to neuropathic pain, where CB2 is up-regulated in macrophages and microglia in response to nerve (and other) injuries, and CB2 agonism reduces macrophage pro-inflamatory (M1) activities. The preclinical efficacy assessment focuses on quantitative measurements, such as measuring how quickly an animal withdraws its paw from heat or pressure insults – tests that can be adapted to human studies and sidestep many of the issues with self-reported pain-scale measurements that are notoriously sensitive to placebo effects.
How are AI and protein language models (pLLMs) supporting antibody discovery efforts?
AI, including pLLMs, is helping in two main ways. First, it is helping us with discovery – that is, boosting signal to noise in our datasets and enabling us to identify hits that more manual bioinformatic analyses would miss.
Second, we are able to extract generalizable sequence-structure information from our datasets. With this predictive capability, we can move from functional antibody “discovery,” where you’re hoping the right functionally active antibody already exists in the starting library and that you can uncover it, into “engineering”, where you have some more control over your outcomes. Functional predictive power also enables us to computationally optimize for multiple criteria simultaneously during antibody sequence selection and generation (function and stability) to shortcut the usual serial “moving target” process.
What’s your roadmap toward the clinic?
We have a development candidate now. Our current roadmap is to develop its application in two disease areas: liver fibrosis and peripheral neuropathy. These are areas of great need but have historically been challenging. Our strategy is to develop these molecules towards the clinic with pharma partners who have a pre-existing strong interest and incentive and expertise in the respective therapeutic area.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
